In our study, we establish a model based on the RSF algorithm for predicting the OS of patients pathologically diagnosed with LACC. In contrast to the Deepsurv and traditional Cox regression models, our RSF model shows better discrimination and calibration ability in predicting 1-, 3-, and 5-year OS. Through the SHAP plot of the RSF model, the SMD was identified as the most significant risk variable, followed by Surgery, FIGO stage, VATD, SATD, SMI, SATI, Hb, BMI, Fever, and Pathology.
In traditional statistical approaches, only theoretically relevant parameters or significant univariate parameters are included into a pre-designed model. In contrast, machine learning allows for the inclusion of numerous variables by discerning patterns from large datasets [23]. RSF, as a novel machine learning method, was constructed based on the decision trees. RSF can make use of internal information cross-validation to avoid overfitting, ensuring the high accuracy of predictive models [24]. Compared to the Cox regression model, which is limited to the proportional hazards assumption, RSF demonstrates its strength of exploring the linear and nonlinear relationships between prognostic factors [18, 25]. Multiple studies have applied the RSF algorithm to predict cancer prognosis and showed a favorable performance, such as for pancreatic cancer, hepatocellular carcinoma, and colorectal cancer, among others [24, 26, 27]. Our research further demonstrates the strength of the RSF algorithm, with AUC values of the RSF model for 1-, 3-, 5-year OS of the test set reaching 0.986, 0.935, and 0.805, respectively.
Incorporating body composition parameters like SMD, SATD, and VATD into prognostic models represents a significant advance over traditional methods that primarily consider clinical and demographic variables. Recent research has increasingly highlighted the prognostic significance of body composition in cancer outcomes. For example, studies suggest that variations in muscle and fat distribution can influence both the efficacy of treatment and survival rates, supporting the notion that body composition metrics are critical predictors of patient outcomes[28, 29]. Our study extends these findings by effectively integrating these metrics into a sophisticated RSF model, offering a more nuanced understanding of prognosis in cervical cancer patients.
In our research, we take advantage of the SHAP plot to explain the output of the developed RSF model. We found that SMD ranks top in variable importance and is an independent risk factor for LACC patients, as determined using multivariate Cox regression analysis. Moreover, we found that patients with a low SMD prior to radiotherapy (SMD<41.1HU) have poorer survival, as confirmed in other studies [11, 30–32]. Sarcopenia and cachexia often occur in cancer patients. SMD, as a radiological characteristic, can adequately reflect ongoing cachexia as well as sarcopenia [33, 34]. Low SMD is indicative of higher adipose tissue production and muscular infiltration, thereby reflecting the poor quality of skeletal muscle [35]. Furthermore, pretreatment SMD is confirmed to be associated with the inflammatory response, chemotherapy-induced toxicity, and postoperative complications in various cancers [34, 36, 37].
BMI is one of the most used prognostic factors. A retrospective study of the association between body fat percentage (BFP, BFP = 1.2 [BMI] + 0.23 [age] − 10.8 [sex] − 5.4) and cervical cancer risk was carried out among patients with either a normal weight or overweight, with the finding that the risk increases with BFP ≥ 45%[38]. In contrast, Lee et al. [39] found that a low value of BMI is associated with a higher risk of grade ≥ 3 late gastrointestinal toxicity in patients with locally advanced cervical cancer treated using definitive IMRT (intensity-modulated radiotherapy). These contradictory findings suggest a potential obesity paradox[40, 41]. Although BMI plays an important role in characterizing body composition, it does not take into account the proportion of tissue components, and patients with the exact same BMI can have significantly different body composition and, therefore, different outcomes. As a result, more useful and specific body composition metrics, such as skeletal muscle mass, VAT, and SAT were recommended [42].
The impact of VAT on cervical cancer is currently unclear. A rapid decline in VAT over 1 month was found to be closely associated with poorer survival in unresectable advanced pancreatic cancer [14]. Increased visceral fat is associated with shortened survival in lymph-node-positive patients with pancreatic adenocarcinoma after pancreatoduodenectomy [43]. VAT modulates the cellular response to radiation in esophageal adenocarcinoma while sensitivity to radiation in noncancer patients is also increased, indicating that the factors driving this effect are not present as a consequence of malignancy but, rather, may be driven by the inflammatory nature of the visceral fat tissue itself [44]. Emerging research suggests VAT plays a key role in modulating a variety of systemic and endocrine effects, as it contains a greater number of matured adipocytes capable of inducing macrophage infiltration and activation [45]. Activated macrophages are important sources of diverse inflammatory cytokines, including tumor necrosis factor α, interleukin-6, and adiponectin, which promote tumor growth [46].
Studies about the impact of SAT on cervical cancer prognosis are still rare. In our study, we found that SATD is an independent risk factor for LACC patients based on multivariate Cox regression analysis, and patients with a higher SATD had poorer OS (P<0.001). Furthermore, the SHAP plot indicates a negative correlation between SATD and OS rate. Similar to our findings, Pellegrini et al. [47] found increased SATD is associated with poorer cancer-specific survival in rectal cancer; Patrick T. Bradshaw et al. [48] found that elevated SAT was associated with worse survival among the women with non-metastatic breast cancer. The associated underlying mechanism may be attributed to the shrinkage, fibrosis, or inflammation of adipose tissue [49, 50]. Moreover, Mangge et al. [51] found that telomere length is inversely correlated with SAT thickness at the upper abdomen, lower abdomen, lower back, and hip, possibly as a result of the direct damaging effects of adipose tissue on telomeres and shelterin proteins due to obesity-induced conditions, such as oxidative stress, inflammation, and physical inactivity. This association of telomere shortening with aging and mortality [52] may be the basis of the link between SATD and poor prognosis. In addition, we used scans at L3 vertebral for our analysis, which mainly captures SAT in the trunk, especially in the abdomen. One possible explanation is that SAT in the abdomen appears to have a different systemic metabolic effect than SAT in the gluteofemoral region [53]. Abdominal SAT can be regarded as consisting of two anatomically and functionally distinct compartments: the superficial SAT (sSAT) and the deep SAT (dSAT) compartments, which are separated by a fascial plane and can be recognized by CT [54, 55]. There seems to be a gradient toward less organization and more vascularization of the adipose tissue depots proceeding from SSAT to DSAT and then to VAT [53]. Abdominal sSAT is metabolically more similar to gluteofemoral SAT than dSAT or VAT as it has less potential for insulin resistance than other compartments [56]. Conversely, researchers have found that dSAT has systemic effects similar to VAT[57]. Thus, SAT around the abdomen may influence tumor promotion and cancer survival through mechanisms similar to those related to VAT.
There are some limitations of this study. It is a retrospective study with a small sample size; thus, bias could affect the results. In addition, we did not conduct further stratified analyses of patients’ weight nor analyze changes in body composition before and after treatment. The impact of changes in these parameters on the outcomes of cervical cancer patients treated with radiotherapy should be further assessed in future through studies using large sample sizes.